How does cost compare at 1M, 10M, and 100M vectors?

At 1M vectors: Pinecone Starter $70/month, Qdrant Cloud ~$25/month, pgvector on Supabase free tier or ~$25/month, self-hosted Qdrant ~$15/month on a small VM. At 10M vectors: Pinecone Standard $700–1,400/month, Qdrant Cloud ~$120/month, self-hosted Qdrant on a 4-core VM ~$80/month, pgvector requires a high-memory Postgres instance at ~$150–300/month. At 100M vectors: Pinecone Enterprise pricing (expect $5,000+/month), self-hosted Qdrant on a dedicated server ~$500–800/month. pgvector at this scale requires significant Postgres expertise and hardware.

Which vector databases support hybrid search?

Weaviate has the most mature hybrid search. It combines dense vector search with BM25 keyword search and merges results with a configurable fusion algorithm. Qdrant added sparse vector support, which enables hybrid search when you also store BM25 or SPLADE sparse vectors. Pinecone supports sparse-dense hybrid with its own sparse index. pgvector alone doesn't do hybrid search, but you can add pg_trgm or full-text search from Postgres and merge results in your application layer.

How hard is it to migrate between vector database providers?

Migrating is mostly a matter of re-inserting your vectors, the embedding model doesn't change, so you're just moving data. The harder part is updating all your application code to use the new client and query format. Pinecone to Qdrant is a common migration path; the LangChain and LlamaIndex integrations abstract most of the difference. The main cost is re-inserting millions of vectors and running validation to confirm recall quality is equivalent.

When does the embedding model matter more than the database choice?

Almost always. The quality of your retrieval is driven 80% by the embedding model and chunking strategy, and 20% by the database. A well-tuned chunking strategy with a domain-appropriate embedding model on pgvector will outperform lazy chunking with a generic model on Pinecone. Choose your embedding model first: text-embedding-3-large for general English, a fine-tuned model for domain-specific retrieval, a multilingual model if you have multiple languages. Then choose the database based on scale and infrastructure needs.

Stack Guide · Vector Databases

Which vector database should you use? It depends on three things.

Pinecone, Qdrant, pgvector, Weaviate. They all store and query vector embeddings. The differences that matter in production are cost, operational complexity, and how well they handle your specific scale. Getting this wrong means either overpaying for simplicity or taking on infra work you didn't budget for.

These are the three factors that should drive the decision, a clear comparison of each option, and our default recommendation for most teams.

RAG development services All stacks →

Building a RAG system?

The three factors that drive the decision.

Before comparing options, be clear on these three variables. They determine which trade-offs matter for you.

Scale: how many vectors are you storing?

Under 500k vectors, almost any option works. Between 500k and 5M, you need to think about index type and query latency. Above 10M, infrastructure complexity and cost become the primary constraints. Most production RAG systems land in the 100k–2M range. That's the sweet spot where pgvector and Qdrant are compelling alternatives to Pinecone.

Operational complexity tolerance

Some teams want a managed service they never think about. Others are comfortable running a Docker container and managing their own infrastructure. Pinecone is the former. Qdrant self-hosted is the latter. Qdrant Cloud and pgvector on RDS or Supabase land in the middle.

Cost sensitivity

Pinecone's pricing is per index and per query, which adds up fast at scale. At 10M vectors with moderate query volume, Pinecone can run $500–2,000/month. Qdrant self-hosted on a single VM costs $50–100/month for the same workload. pgvector on an existing Postgres instance is essentially free if you're already paying for the database.

Pinecone.

Managed, opinionated, and the easiest path from nothing to working vector search. You pay for that simplicity. It's the most expensive option at scale.

What works

Fully managed, no infra to run
Fastest path from zero to working queries
Reliable at any scale, handles spikes without tuning
Good official SDK support for Python, Node, Go

Trade-offs

Most expensive option at scale
Vendor lock-in: migrating out requires re-embedding
Limited filtering compared to Qdrant's payload filters
No on-premise option

Right for

Prototypes, teams that prioritize speed-to-production over cost, and organizations that don't want to manage vector infrastructure at any scale.

Qdrant.

Open-source, self-host or cloud. Better price-to-performance than Pinecone at scale. The right default when cost matters and you can tolerate a small amount of infrastructure work.

What works

Best price-to-performance ratio at scale
Rich filtering on payload metadata
Self-host or use Qdrant Cloud
On-premise option for data residency requirements
Fast updates, handles frequent inserts better than Pinecone

Trade-offs

More setup than Pinecone if self-hosting
Smaller ecosystem than LangChain/LlamaIndex Pinecone integrations
Documentation is good but thinner than Pinecone's

Right for

Cost-sensitive teams, on-premise requirements, use cases with heavy metadata filtering, and production systems above 2M vectors.

pgvector.

A Postgres extension that adds vector storage and similarity search. If you're already on Postgres, this is the lowest-friction option for moderate-scale vector search.

What works

No additional infrastructure if you're already on Postgres
Full SQL for metadata filtering, no new query language
ACID transactions alongside vector operations
Works on Supabase, RDS, Neon, any managed Postgres

Trade-offs

Slower at high vector counts than dedicated vector DBs
Index types (IVFFlat, HNSW) require tuning for large collections
Not the right choice above ~5M vectors without careful planning

Right for

Teams already running Postgres, vector counts under 1M, use cases where the vectors are tightly coupled to relational data, and situations where adding another database is not worth it.

Weaviate.

GraphQL interface, native hybrid search, built-in multi-tenancy. The most feature-rich option, and the most complex to configure well.

What works

Built-in hybrid search (vector + keyword BM25)
GraphQL query interface
Multi-tenancy support built in
Module system for embedding and reranking

Trade-offs

More complex to configure than alternatives
GraphQL adds a learning curve for teams used to REST
Smaller community than Pinecone or pgvector

Right for

Search-heavy applications that need hybrid retrieval (semantic + keyword), teams comfortable with GraphQL, and multi-tenant SaaS applications.

What we actually recommend.

Start with pgvector if Postgres is already in your stack and your document count is under 1M. Zero additional infrastructure, transactions alongside vector operations, and familiar SQL for filtering.

Move to Qdrant when you need scale beyond what pgvector handles well, when you need richer metadata filtering, or when cost at scale is a constraint. Self-host for the lowest cost; use Qdrant Cloud if you want managed without the Pinecone price tag.

Use Pinecone when you need the fastest path to production and don't want to think about vector infrastructure at all. Accept that you'll pay for that simplicity as you scale.

Consider Weaviate only if you specifically need hybrid search as a first-class feature or have a multi-tenant SaaS use case where its built-in tenancy model is a genuine fit.

Frequently asked questions.

How does cost compare at 1M, 10M, and 100M vectors?: At 1M vectors: Pinecone Starter $70/month, Qdrant Cloud ~$25/month, pgvector on Supabase free tier or ~$25/month, self-hosted Qdrant ~$15/month on a small VM. At 10M vectors: Pinecone Standard $700–1,400/month, Qdrant Cloud ~$120/month, self-hosted Qdrant on a 4-core VM ~$80/month, pgvector requires a high-memory Postgres instance at ~$150–300/month. At 100M vectors: Pinecone Enterprise pricing (expect $5,000+/month), self-hosted Qdrant on a dedicated server ~$500–800/month. pgvector at this scale requires significant Postgres expertise and hardware.
Which vector databases support hybrid search?: Weaviate has the most mature hybrid search. It combines dense vector search with BM25 keyword search and merges results with a configurable fusion algorithm. Qdrant added sparse vector support, which enables hybrid search when you also store BM25 or SPLADE sparse vectors. Pinecone supports sparse-dense hybrid with its own sparse index. pgvector alone doesn't do hybrid search, but you can add pg_trgm or full-text search from Postgres and merge results in your application layer.
How hard is it to migrate between vector database providers?: Migrating is mostly a matter of re-inserting your vectors, the embedding model doesn't change, so you're just moving data. The harder part is updating all your application code to use the new client and query format. Pinecone to Qdrant is a common migration path; the LangChain and LlamaIndex integrations abstract most of the difference. The main cost is re-inserting millions of vectors and running validation to confirm recall quality is equivalent.
When does the embedding model matter more than the database choice?: Almost always. The quality of your retrieval is driven 80% by the embedding model and chunking strategy, and 20% by the database. A well-tuned chunking strategy with a domain-appropriate embedding model on pgvector will outperform lazy chunking with a generic model on Pinecone. Choose your embedding model first: text-embedding-3-large for general English, a fine-tuned model for domain-specific retrieval, a multilingual model if you have multiple languages. Then choose the database based on scale and infrastructure needs.

Building a RAG system and not sure which database to use?

Tell us your document volume, query patterns, and infrastructure preferences. We can recommend the right option and build the full RAG pipeline around it.